1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vsigmoid/sse-rr2-lut64-p2-div.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <emmintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vunary.h>
16
17
18 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_64[64];
19
xnn_f32_vsigmoid_ukernel__sse2_rr2_lut64_p2_div_x24(size_t n,const float * x,float * y,const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f32_vsigmoid_ukernel__sse2_rr2_lut64_p2_div_x24(
21 size_t n,
22 const float* x,
23 float* y,
24 const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25 {
26 assert(n % sizeof(float) == 0);
27
28 const __m128 vsign_mask = _mm_load_ps(params->sse2_rr2_lut64_p2.sign_mask);
29 const __m128 vmagic_bias = _mm_load_ps(params->sse2_rr2_lut64_p2.magic_bias);
30 const __m128 vlog2e = _mm_load_ps(params->sse2_rr2_lut64_p2.log2e);
31 const __m128i vindex_mask = _mm_load_si128((const __m128i*) params->sse2_rr2_lut64_p2.index_mask);
32 const __m128 vminus_ln2_hi = _mm_load_ps(params->sse2_rr2_lut64_p2.minus_ln2_hi);
33 const __m128 vminus_ln2_lo = _mm_load_ps(params->sse2_rr2_lut64_p2.minus_ln2_lo);
34 const __m128 vc2 = _mm_load_ps(params->sse2_rr2_lut64_p2.c2);
35 const __m128 vone = _mm_load_ps(params->sse2_rr2_lut64_p2.one);
36 const __m128 vdenorm_cutoff = _mm_load_ps(params->sse2_rr2_lut64_p2.denorm_cutoff);
37
38 for (; n >= 24 * sizeof(float); n -= 24 * sizeof(float)) {
39 const __m128 vx0123 = _mm_loadu_ps(x);
40 const __m128 vx4567 = _mm_loadu_ps(x + 4);
41 const __m128 vx89AB = _mm_loadu_ps(x + 8);
42 const __m128 vxCDEF = _mm_loadu_ps(x + 12);
43 const __m128 vxGHIJ = _mm_loadu_ps(x + 16);
44 const __m128 vxKLMN = _mm_loadu_ps(x + 20);
45 x += 24;
46
47 const __m128 vz0123 = _mm_or_ps(vx0123, vsign_mask);
48 const __m128 vz4567 = _mm_or_ps(vx4567, vsign_mask);
49 const __m128 vz89AB = _mm_or_ps(vx89AB, vsign_mask);
50 const __m128 vzCDEF = _mm_or_ps(vxCDEF, vsign_mask);
51 const __m128 vzGHIJ = _mm_or_ps(vxGHIJ, vsign_mask);
52 const __m128 vzKLMN = _mm_or_ps(vxKLMN, vsign_mask);
53
54 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias);
55 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias);
56 __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vz89AB, vlog2e), vmagic_bias);
57 __m128 vnCDEF = _mm_add_ps(_mm_mul_ps(vzCDEF, vlog2e), vmagic_bias);
58 __m128 vnGHIJ = _mm_add_ps(_mm_mul_ps(vzGHIJ, vlog2e), vmagic_bias);
59 __m128 vnKLMN = _mm_add_ps(_mm_mul_ps(vzKLMN, vlog2e), vmagic_bias);
60
61 const __m128i ve0123 = _mm_slli_epi32(_mm_castps_si128(vn0123), 17);
62 const __m128i ve4567 = _mm_slli_epi32(_mm_castps_si128(vn4567), 17);
63 const __m128i ve89AB = _mm_slli_epi32(_mm_castps_si128(vn89AB), 17);
64 const __m128i veCDEF = _mm_slli_epi32(_mm_castps_si128(vnCDEF), 17);
65 const __m128i veGHIJ = _mm_slli_epi32(_mm_castps_si128(vnGHIJ), 17);
66 const __m128i veKLMN = _mm_slli_epi32(_mm_castps_si128(vnKLMN), 17);
67
68 const __m128i vidx0123 = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn0123), vindex_mask), 2);
69 const __m128i vidx4567 = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn4567), vindex_mask), 2);
70 const __m128i vidx89AB = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn89AB), vindex_mask), 2);
71 const __m128i vidxCDEF = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vnCDEF), vindex_mask), 2);
72 const __m128i vidxGHIJ = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vnGHIJ), vindex_mask), 2);
73 const __m128i vidxKLMN = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vnKLMN), vindex_mask), 2);
74
75 #if XNN_ARCH_X86_64
76 const uint64_t vidx01 = (uint64_t) _mm_cvtsi128_si64(vidx0123);
77 const uint64_t vidx23 = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx0123, vidx0123));
78 const __m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx01)));
79 const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx23)));
80 const __m128i vl1 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx01 >> 32))));
81 const __m128i vl01 = _mm_unpacklo_epi32(vl0, vl1);
82 const __m128i vl3 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx23 >> 32))));
83 const __m128i vl23 = _mm_unpacklo_epi32(vl2, vl3);
84 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
85 const uint64_t vidx45 = (uint64_t) _mm_cvtsi128_si64(vidx4567);
86 const uint64_t vidx67 = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx4567, vidx4567));
87 const __m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx45)));
88 const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx67)));
89 const __m128i vl5 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx45 >> 32))));
90 const __m128i vl45 = _mm_unpacklo_epi32(vl4, vl5);
91 const __m128i vl7 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx67 >> 32))));
92 const __m128i vl67 = _mm_unpacklo_epi32(vl6, vl7);
93 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
94 const uint64_t vidx89 = (uint64_t) _mm_cvtsi128_si64(vidx89AB);
95 const uint64_t vidxAB = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx89AB, vidx89AB));
96 const __m128i vl8 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx89)));
97 const __m128i vlA = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxAB)));
98 const __m128i vl9 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx89 >> 32))));
99 const __m128i vl89 = _mm_unpacklo_epi32(vl8, vl9);
100 const __m128i vlB = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxAB >> 32))));
101 const __m128i vlAB = _mm_unpacklo_epi32(vlA, vlB);
102 const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB);
103 const uint64_t vidxCD = (uint64_t) _mm_cvtsi128_si64(vidxCDEF);
104 const uint64_t vidxEF = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidxCDEF, vidxCDEF));
105 const __m128i vlC = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxCD)));
106 const __m128i vlE = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxEF)));
107 const __m128i vlD = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxCD >> 32))));
108 const __m128i vlCD = _mm_unpacklo_epi32(vlC, vlD);
109 const __m128i vlF = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxEF >> 32))));
110 const __m128i vlEF = _mm_unpacklo_epi32(vlE, vlF);
111 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF);
112 const uint64_t vidxGH = (uint64_t) _mm_cvtsi128_si64(vidxGHIJ);
113 const uint64_t vidxIJ = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidxGHIJ, vidxGHIJ));
114 const __m128i vlG = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxGH)));
115 const __m128i vlI = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxIJ)));
116 const __m128i vlH = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxGH >> 32))));
117 const __m128i vlGH = _mm_unpacklo_epi32(vlG, vlH);
118 const __m128i vlJ = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxIJ >> 32))));
119 const __m128i vlIJ = _mm_unpacklo_epi32(vlI, vlJ);
120 const __m128i vlGHIJ = _mm_unpacklo_epi64(vlGH, vlIJ);
121 const uint64_t vidxKL = (uint64_t) _mm_cvtsi128_si64(vidxKLMN);
122 const uint64_t vidxMN = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidxKLMN, vidxKLMN));
123 const __m128i vlK = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxKL)));
124 const __m128i vlM = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxMN)));
125 const __m128i vlL = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxKL >> 32))));
126 const __m128i vlKL = _mm_unpacklo_epi32(vlK, vlL);
127 const __m128i vlN = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxMN >> 32))));
128 const __m128i vlMN = _mm_unpacklo_epi32(vlM, vlN);
129 const __m128i vlKLMN = _mm_unpacklo_epi64(vlKL, vlMN);
130 #else // !XNN_ARCH_X86_64
131 const uint32_t vidx0 = (uint32_t) _mm_cvtsi128_si32(vidx0123);
132 const uint32_t vidx1 = (uint32_t) _mm_extract_epi16(vidx0123, 2);
133 const uint32_t vidx2 = (uint32_t) _mm_extract_epi16(vidx0123, 4);
134 const uint32_t vidx3 = (uint32_t) _mm_extract_epi16(vidx0123, 6);
135 const __m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx0)));
136 const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx2)));
137 const __m128i vl1 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx1)));
138 const __m128i vl01 = _mm_unpacklo_epi32(vl0, vl1);
139 const __m128i vl3 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx3)));
140 const __m128i vl23 = _mm_unpacklo_epi32(vl2, vl3);
141 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
142 const uint32_t vidx4 = (uint32_t) _mm_cvtsi128_si32(vidx4567);
143 const uint32_t vidx5 = (uint32_t) _mm_extract_epi16(vidx4567, 2);
144 const uint32_t vidx6 = (uint32_t) _mm_extract_epi16(vidx4567, 4);
145 const uint32_t vidx7 = (uint32_t) _mm_extract_epi16(vidx4567, 6);
146 const __m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx4)));
147 const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx6)));
148 const __m128i vl5 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx5)));
149 const __m128i vl45 = _mm_unpacklo_epi32(vl4, vl5);
150 const __m128i vl7 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx7)));
151 const __m128i vl67 = _mm_unpacklo_epi32(vl6, vl7);
152 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
153 const uint32_t vidx8 = (uint32_t) _mm_cvtsi128_si32(vidx89AB);
154 const uint32_t vidx9 = (uint32_t) _mm_extract_epi16(vidx89AB, 2);
155 const uint32_t vidxA = (uint32_t) _mm_extract_epi16(vidx89AB, 4);
156 const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6);
157 const __m128i vl8 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx8)));
158 const __m128i vlA = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxA)));
159 const __m128i vl9 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx9)));
160 const __m128i vl89 = _mm_unpacklo_epi32(vl8, vl9);
161 const __m128i vlB = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxB)));
162 const __m128i vlAB = _mm_unpacklo_epi32(vlA, vlB);
163 const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB);
164 const uint32_t vidxC = (uint32_t) _mm_cvtsi128_si32(vidxCDEF);
165 const uint32_t vidxD = (uint32_t) _mm_extract_epi16(vidxCDEF, 2);
166 const uint32_t vidxE = (uint32_t) _mm_extract_epi16(vidxCDEF, 4);
167 const uint32_t vidxF = (uint32_t) _mm_extract_epi16(vidxCDEF, 6);
168 const __m128i vlC = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxC)));
169 const __m128i vlE = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxE)));
170 const __m128i vlD = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxD)));
171 const __m128i vlCD = _mm_unpacklo_epi32(vlC, vlD);
172 const __m128i vlF = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxF)));
173 const __m128i vlEF = _mm_unpacklo_epi32(vlE, vlF);
174 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF);
175 const uint32_t vidxG = (uint32_t) _mm_cvtsi128_si32(vidxGHIJ);
176 const uint32_t vidxH = (uint32_t) _mm_extract_epi16(vidxGHIJ, 2);
177 const uint32_t vidxI = (uint32_t) _mm_extract_epi16(vidxGHIJ, 4);
178 const uint32_t vidxJ = (uint32_t) _mm_extract_epi16(vidxGHIJ, 6);
179 const __m128i vlG = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxG)));
180 const __m128i vlI = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxI)));
181 const __m128i vlH = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxH)));
182 const __m128i vlGH = _mm_unpacklo_epi32(vlG, vlH);
183 const __m128i vlJ = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxJ)));
184 const __m128i vlIJ = _mm_unpacklo_epi32(vlI, vlJ);
185 const __m128i vlGHIJ = _mm_unpacklo_epi64(vlGH, vlIJ);
186 const uint32_t vidxK = (uint32_t) _mm_cvtsi128_si32(vidxKLMN);
187 const uint32_t vidxL = (uint32_t) _mm_extract_epi16(vidxKLMN, 2);
188 const uint32_t vidxM = (uint32_t) _mm_extract_epi16(vidxKLMN, 4);
189 const uint32_t vidxN = (uint32_t) _mm_extract_epi16(vidxKLMN, 6);
190 const __m128i vlK = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxK)));
191 const __m128i vlM = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxM)));
192 const __m128i vlL = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxL)));
193 const __m128i vlKL = _mm_unpacklo_epi32(vlK, vlL);
194 const __m128i vlN = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxN)));
195 const __m128i vlMN = _mm_unpacklo_epi32(vlM, vlN);
196 const __m128i vlKLMN = _mm_unpacklo_epi64(vlKL, vlMN);
197 #endif // XNN_ARCH_X86_64
198
199 const __m128 vs0123 = _mm_castsi128_ps(_mm_add_epi32(vl0123, ve0123));
200 const __m128 vs4567 = _mm_castsi128_ps(_mm_add_epi32(vl4567, ve4567));
201 const __m128 vs89AB = _mm_castsi128_ps(_mm_add_epi32(vl89AB, ve89AB));
202 const __m128 vsCDEF = _mm_castsi128_ps(_mm_add_epi32(vlCDEF, veCDEF));
203 const __m128 vsGHIJ = _mm_castsi128_ps(_mm_add_epi32(vlGHIJ, veGHIJ));
204 const __m128 vsKLMN = _mm_castsi128_ps(_mm_add_epi32(vlKLMN, veKLMN));
205
206 vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
207 vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
208 vn89AB = _mm_sub_ps(vn89AB, vmagic_bias);
209 vnCDEF = _mm_sub_ps(vnCDEF, vmagic_bias);
210 vnGHIJ = _mm_sub_ps(vnGHIJ, vmagic_bias);
211 vnKLMN = _mm_sub_ps(vnKLMN, vmagic_bias);
212
213 __m128 vt0123 = _mm_add_ps(vz0123, _mm_mul_ps(vn0123, vminus_ln2_hi));
214 __m128 vt4567 = _mm_add_ps(vz4567, _mm_mul_ps(vn4567, vminus_ln2_hi));
215 __m128 vt89AB = _mm_add_ps(vz89AB, _mm_mul_ps(vn89AB, vminus_ln2_hi));
216 __m128 vtCDEF = _mm_add_ps(vzCDEF, _mm_mul_ps(vnCDEF, vminus_ln2_hi));
217 __m128 vtGHIJ = _mm_add_ps(vzGHIJ, _mm_mul_ps(vnGHIJ, vminus_ln2_hi));
218 __m128 vtKLMN = _mm_add_ps(vzKLMN, _mm_mul_ps(vnKLMN, vminus_ln2_hi));
219
220 vt0123 = _mm_add_ps(vt0123, _mm_mul_ps(vn0123, vminus_ln2_lo));
221 vt4567 = _mm_add_ps(vt4567, _mm_mul_ps(vn4567, vminus_ln2_lo));
222 vt89AB = _mm_add_ps(vt89AB, _mm_mul_ps(vn89AB, vminus_ln2_lo));
223 vtCDEF = _mm_add_ps(vtCDEF, _mm_mul_ps(vnCDEF, vminus_ln2_lo));
224 vtGHIJ = _mm_add_ps(vtGHIJ, _mm_mul_ps(vnGHIJ, vminus_ln2_lo));
225 vtKLMN = _mm_add_ps(vtKLMN, _mm_mul_ps(vnKLMN, vminus_ln2_lo));
226
227 __m128 vp0123 = _mm_mul_ps(vt0123, vc2);
228 __m128 vp4567 = _mm_mul_ps(vt4567, vc2);
229 __m128 vp89AB = _mm_mul_ps(vt89AB, vc2);
230 __m128 vpCDEF = _mm_mul_ps(vtCDEF, vc2);
231 __m128 vpGHIJ = _mm_mul_ps(vtGHIJ, vc2);
232 __m128 vpKLMN = _mm_mul_ps(vtKLMN, vc2);
233
234 vp0123 = _mm_add_ps(vt0123, _mm_mul_ps(vp0123, vt0123));
235 vp4567 = _mm_add_ps(vt4567, _mm_mul_ps(vp4567, vt4567));
236 vp89AB = _mm_add_ps(vt89AB, _mm_mul_ps(vp89AB, vt89AB));
237 vpCDEF = _mm_add_ps(vtCDEF, _mm_mul_ps(vpCDEF, vtCDEF));
238 vpGHIJ = _mm_add_ps(vtGHIJ, _mm_mul_ps(vpGHIJ, vtGHIJ));
239 vpKLMN = _mm_add_ps(vtKLMN, _mm_mul_ps(vpKLMN, vtKLMN));
240
241 const __m128 vy0123 = _mm_add_ps(vs0123, _mm_mul_ps(vs0123, vp0123));
242 const __m128 vy4567 = _mm_add_ps(vs4567, _mm_mul_ps(vs4567, vp4567));
243 const __m128 vy89AB = _mm_add_ps(vs89AB, _mm_mul_ps(vs89AB, vp89AB));
244 const __m128 vyCDEF = _mm_add_ps(vsCDEF, _mm_mul_ps(vsCDEF, vpCDEF));
245 const __m128 vyGHIJ = _mm_add_ps(vsGHIJ, _mm_mul_ps(vsGHIJ, vpGHIJ));
246 const __m128 vyKLMN = _mm_add_ps(vsKLMN, _mm_mul_ps(vsKLMN, vpKLMN));
247
248 __m128 vf0123 = _mm_div_ps(vy0123, _mm_add_ps(vy0123, vone));
249 __m128 vf4567 = _mm_div_ps(vy4567, _mm_add_ps(vy4567, vone));
250 __m128 vf89AB = _mm_div_ps(vy89AB, _mm_add_ps(vy89AB, vone));
251 __m128 vfCDEF = _mm_div_ps(vyCDEF, _mm_add_ps(vyCDEF, vone));
252 __m128 vfGHIJ = _mm_div_ps(vyGHIJ, _mm_add_ps(vyGHIJ, vone));
253 __m128 vfKLMN = _mm_div_ps(vyKLMN, _mm_add_ps(vyKLMN, vone));
254
255 vf0123 = _mm_andnot_ps(_mm_cmplt_ps(vz0123, vdenorm_cutoff), vf0123);
256 vf4567 = _mm_andnot_ps(_mm_cmplt_ps(vz4567, vdenorm_cutoff), vf4567);
257 vf89AB = _mm_andnot_ps(_mm_cmplt_ps(vz89AB, vdenorm_cutoff), vf89AB);
258 vfCDEF = _mm_andnot_ps(_mm_cmplt_ps(vzCDEF, vdenorm_cutoff), vfCDEF);
259 vfGHIJ = _mm_andnot_ps(_mm_cmplt_ps(vzGHIJ, vdenorm_cutoff), vfGHIJ);
260 vfKLMN = _mm_andnot_ps(_mm_cmplt_ps(vzKLMN, vdenorm_cutoff), vfKLMN);
261
262 const __m128 vm0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx0123)));
263 const __m128 vm4567 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx4567)));
264 const __m128 vm89AB = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx89AB)));
265 const __m128 vmCDEF = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxCDEF)));
266 const __m128 vmGHIJ = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxGHIJ)));
267 const __m128 vmKLMN = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxKLMN)));
268
269 vf0123 = _mm_or_ps(_mm_and_ps(vf0123, vm0123), _mm_andnot_ps(vm0123, _mm_sub_ps(vone, vf0123)));
270 vf4567 = _mm_or_ps(_mm_and_ps(vf4567, vm4567), _mm_andnot_ps(vm4567, _mm_sub_ps(vone, vf4567)));
271 vf89AB = _mm_or_ps(_mm_and_ps(vf89AB, vm89AB), _mm_andnot_ps(vm89AB, _mm_sub_ps(vone, vf89AB)));
272 vfCDEF = _mm_or_ps(_mm_and_ps(vfCDEF, vmCDEF), _mm_andnot_ps(vmCDEF, _mm_sub_ps(vone, vfCDEF)));
273 vfGHIJ = _mm_or_ps(_mm_and_ps(vfGHIJ, vmGHIJ), _mm_andnot_ps(vmGHIJ, _mm_sub_ps(vone, vfGHIJ)));
274 vfKLMN = _mm_or_ps(_mm_and_ps(vfKLMN, vmKLMN), _mm_andnot_ps(vmKLMN, _mm_sub_ps(vone, vfKLMN)));
275
276 _mm_storeu_ps(y, vf0123);
277 _mm_storeu_ps(y + 4, vf4567);
278 _mm_storeu_ps(y + 8, vf89AB);
279 _mm_storeu_ps(y + 12, vfCDEF);
280 _mm_storeu_ps(y + 16, vfGHIJ);
281 _mm_storeu_ps(y + 20, vfKLMN);
282 y += 24;
283 }
284 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
285 const __m128 vx = _mm_loadu_ps(x);
286 x += 4;
287
288 const __m128 vz = _mm_or_ps(vx, vsign_mask);
289
290 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
291 const __m128i ve = _mm_slli_epi32(_mm_castps_si128(vn), 17);
292
293 const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
294 #if XNN_ARCH_X86_64
295 const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
296 const uint64_t vidx_hi = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx, vidx));
297 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo)));
298 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi)));
299 const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_lo >> 32))));
300 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
301 const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_hi >> 32))));
302 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
303 #else // !XNN_ARCH_X86_64
304 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_cvtsi128_si32(vidx))));
305 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 4))));
306 const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 2))));
307 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
308 const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 6))));
309 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
310 #endif // XNN_ARCH_X86_64
311 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
312
313 const __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ve));
314 vn = _mm_sub_ps(vn, vmagic_bias);
315
316 __m128 vt = _mm_add_ps(vz, _mm_mul_ps(vn, vminus_ln2_hi));
317 vt = _mm_add_ps(vt, _mm_mul_ps(vn, vminus_ln2_lo));
318
319 __m128 vp = _mm_mul_ps(vt, vc2);
320 vp = _mm_add_ps(vt, _mm_mul_ps(vp, vt));
321
322 const __m128 vy = _mm_add_ps(vs, _mm_mul_ps(vs, vp));
323
324 __m128 vf = _mm_div_ps(vy, _mm_add_ps(vy, vone));
325 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
326 const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
327 vf = _mm_or_ps(_mm_and_ps(vf, vm), _mm_andnot_ps(vm, _mm_sub_ps(vone, vf)));
328
329 _mm_storeu_ps(y, vf);
330 y += 4;
331 }
332 if XNN_UNLIKELY(n != 0) {
333 const __m128 vx = _mm_loadu_ps(x);
334
335 const __m128 vz = _mm_or_ps(vx, vsign_mask);
336
337 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
338 const __m128i ve = _mm_slli_epi32(_mm_castps_si128(vn), 17);
339
340 const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
341 #if XNN_ARCH_X86_64
342 const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
343 const uint64_t vidx_hi = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx, vidx));
344 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo)));
345 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi)));
346 const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_lo >> 32))));
347 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
348 const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_hi >> 32))));
349 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
350 #else // !XNN_ARCH_X86_64
351 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_cvtsi128_si32(vidx))));
352 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 4))));
353 const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 2))));
354 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
355 const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 6))));
356 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
357 #endif // XNN_ARCH_X86_64
358 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
359
360 const __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ve));
361 vn = _mm_sub_ps(vn, vmagic_bias);
362
363 __m128 vt = _mm_add_ps(vz, _mm_mul_ps(vn, vminus_ln2_hi));
364 vt = _mm_add_ps(vt, _mm_mul_ps(vn, vminus_ln2_lo));
365
366 __m128 vp = _mm_mul_ps(vt, vc2);
367 vp = _mm_add_ps(vt, _mm_mul_ps(vp, vt));
368
369 const __m128 vy = _mm_add_ps(vs, _mm_mul_ps(vs, vp));
370
371 __m128 vf = _mm_div_ps(vy, _mm_add_ps(vy, vone));
372 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
373 const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
374 vf = _mm_or_ps(_mm_and_ps(vf, vm), _mm_andnot_ps(vm, _mm_sub_ps(vone, vf)));
375
376 if (n & (2 * sizeof(float))) {
377 _mm_storel_pi((__m64*) y, vf);
378 vf = _mm_movehl_ps(vf, vf);
379 y += 2;
380 }
381 if (n & (1 * sizeof(float))) {
382 _mm_store_ss(y, vf);
383 }
384 }
385 }
386